AI Image Normality Detection Using Sliding AUROC Windows
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing artificial intelligence models using unsupervised learning incorrectly recognize normal images as abnormal due to normalization processes that amplify minor differences or distort original features.
Innovation Solution
Anomaly detection model is trained to identify abnormal regions by generating normal and abnormal learning datasets, performing image data augmentation, and using a sliding AUROC method to analyze inference results in small inspection windows, calculating AUROC values, and comparing them with predefined thresholds to determine image normality.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If normalization is applied to standardize pixel values, then the model learns more consistently, but minute changes are overemphasized causing normal images to be misrecognized as abnormal
Solution Approach 1:
The image is divided into multiple sub-regions (e.g., 8 sub-regions from 2x2 grid) and each region is analyzed separately using sliding windows. This segmentation allows the system to identify localized abnormal regions without being misled by normalization-induced distortions in the entire image, thereby maintaining learning consistency while improving normality determination accuracy.
Solution Approach 2:
Different inspection strategies are applied to different regions of the image. The system calculates AUROC values for each sub-region and identifies abnormal regions based on local characteristics rather than applying a uniform threshold to the entire image. This local quality approach prevents normal regions from being misclassified due to normalization effects.
2Device complexity
If the entire image is analyzed as a single region, then the processing is simpler, but the location of abnormal regions cannot be precisely identified
Solution Approach 1:
The image is divided into multiple sub-regions and further analyzed using sliding windows of various sizes. This segmentation enables precise identification of abnormal region locations while maintaining manageable complexity through systematic processing of divided regions rather than attempting to analyze the entire image at once.
Solution Approach 2:
The analysis is performed across multiple dimensions by using sliding windows of different sizes (e.g., 10x10, 20x20, 30x30 pixels) and analyzing multiple sub-regions. This multi-dimensional approach provides precise location information for abnormal regions while keeping the complexity manageable through hierarchical analysis.
3Measurement precision
If small inspection windows are used to analyze local regions, then abnormal regions can be precisely identified, but the processing time increases
Solution Approach 1:
The image is pre-divided into sub-regions before applying sliding window analysis. This segmentation reduces the total number of sliding window operations needed compared to analyzing the entire image with small windows, thereby maintaining abnormal region detection precision while reducing processing time through hierarchical processing.
Solution Approach 2:
The system uses multiple window sizes (10x10, 20x20, 30x30) and analyzes only regions that show abnormal characteristics. Rather than exhaustively analyzing every possible position with the smallest window, the system applies partial action by focusing computational resources on suspicious regions identified through coarser analysis first.
Data Source
AI summary
A method and apparatus determine image normality using an artificial intelligence model. A method for determining image normality using an anomaly detection model includes obtaining an inference result of the anomaly detection model. The method further includes determining a step size and a size of an inspection window for identifying abnormal regions. The method also includes calculating an AUROC value based on the inspection window. The method further includes determining whether a region of the inspection window is normal or abnormal by comparing the AUROC value with a predefined threshold. The method also includes determining whether an image is normal or abnormal based on the result of identifying abnormal regions.